Hacker News Reader

Comment by SR2Z | Hacker News Reader

I don't think that anything you said is wrong but I also don't think that memory consumption is going back to the old baseline... well, ever.

Memory is just too useful now that you can use it to drive cars and write code.

View on HN

People are happy to play with AI when the tech companies are burning hundreds of billions of dollars to subsidize it. It remains to be seen who is actually willing to pay for it at the prices required to recoup those insane investments.

Which directly leads to the next big development: all the big players are investing in silicon with "baked-in" models, like [0,1]. Turns out you don't need an expensive general-purpose GPU with heaps of RAM to contain a model when you can make a custom ASIC around one specific model! Why spend a fortune on DRAM / HBM when all you need is some finetuning parameters which are easily stored in on-die SRAM?

[0]: https://www.theregister.com/systems/2026/08/06/amd-acquires-...

[1]: https://thenextweb.com/news/google-frozen-chip-gemini-silico...

> It remains to be seen who is actually willing to pay for it at the prices required to recoup those insane investments.

Everyone. Open models are going to keep prices down. A lot of current models are more than usable. Self-hosting would have been an option if hardware prices get back to sane values.

The insane investments have to do with insane over-valuations, VCs involved and hence media spam on it. Chinese labs for example make do with 1% of the valuation and 1% of the resources.

Prices must raise a lot to make self hosting mainstream or at least fairly popular.

Yesterday a coworker posted on a customer's Slack the specs of a box he is planning to buy to run local models. It's about 5k Euro. I am paying 18 Euro per month for Claude Pro and I'm going through a migration (almost a total rewrite) of a web app from Vue 2 / Vuetify 2 / Vuex to Vue 3 / Vuetify 4 / Pinia. I never hit the 6 hours limit. I could consider running an equivalent model on a 500 Euro machine (about 2 years of Claude Pro) but 5k is 20 years and that box will be obsolete or will have failed beyond repair (no spares) much earlier than that.

The problem is that is speculation and the market barriers are currently too high for losing business to be a threat. The two outcomes are (1) you overbuild, you end being wrong, you go bankrupt and lose everything or (2) you are right, but since you didn't overbuild, you lost out on some revenue, but demand still exists 3 years later and you didn't lose everything.

In any other business choosing (2) would mean someone else swoops in and steals all your business. It doesn't look like this is at all possible for memory fabs.

Well, the Chinese manufacturers are certainly coming, though CXMT seems to be being careful not to rock the boat, at least yet https://www.tomshardware.com/pc-components/dram/chinese-cxmt... .

It makes sense, CXMT takes as much profit as possible, then use that money to build more capacity, ultimately becoming one of the biggest competitor (only?) to the cartel.

As of Q1 2026: 90% market share for the Big Three, 8% CXMT, 2% everyone else. And CXMT was only 3% Q1 2025.

"In any other business choosing (2) would mean someone else swoops in and steals all your business. It doesn't look like this is at all possible for memory fabs."

FWIW - Lots of people seem to believe the "in any other business", but it turns out there are tons and tons of places it is equally untrue, even where there is near zero time or cost barrier to entry.

> Memory is just too useful now that you can use it to drive cars and write code.

You can use the state of the art from 20 years ago for that.

Though I am sure with enough time, someone can find a reason to need ECC.

You can't, the bandwidth is too low (unless you mean hiring a human, in which case it's quite expensive)

Reply on news.ycombinator.com